1. Introduction
The relationship between morphological complexity and lexical access has been a paramount question in the field of psycholinguistics for decades, with an increase in research on morphological processing since the 1970s. However, the vast bulk of research in this domain has been carried out on a handful of languages (e.g., English, Dutch, German) with very similar morphological characteristics: roots frequently of just one or two syllables, few if any inflectional prefixes, and the bulk of morphology being suffixal (Anand et al., 2011; Blasi et al., 2022; Bornkessel-Schlesewsky & Schlesewsky, 2016; Collart, 2024; Kidd & Garcia, 2022; Norcliffe et al., 2015). Similarly, more than half of all morphological processing research has been conducted on suffixation (Cayado & Rastle, 2025).
The present research departs from this approach and examines the processing of two morphophonological patterns of alternation in the Australian Indigenous language Wubuy. Wubuy, also known as Nunggubuyu (Heath, 1984; ISO 639-3 code nuy), is a highly polysynthetic language in which word-internal morphological boundaries exhibit two unusual, linked patterns of morphophonological alternation: phonological hardening and ŋu-insertion, both involving the initial segments of stems. Phonological hardening (Section 1.1) results in morpheme-initial phoneme alternations between continuants and stops under conditions where the preceding morpheme ends in a non-continuant (a nasal or a stop; there are no fricatives in the inventory), as in Example (1) and Example (2), where the verb wult̪a ‘cut’ has an initial segment /w/ following continuants as in (1), but is realised as kult̪a following non-continuants such as the stop /c/ in (2).
(1)
[ŋa-wu-wult̪a + ŋi]
1sg-neut-cut+past.cont
‘I was cutting it (neut)’
(2)
[ŋa-wu-ɻuluc-kult̪a + ŋi]
1sg-neut-bough.shade-cut+past.cont
‘I was cutting a bough shade (neut)’
The process of ŋu-insertion (Section 1.1) involves the appearance of the meaningless string /ŋu/ ‘Ø’ before (underlyingly) stop-initial stems under conditions which are partly phonological and partly morphological, but which partially overlap with the conditions for hardening, so that all environments that condition hardening can also condition ŋu-insertion. “Hardened” verbs like kult̪a, in Example (2), contrast with underlyingly stop-initial inflected verbs such as kucukucaani ‘tickle.past.cont’, in Example (3), where it follows a vowel-final agreement prefix. Verbs which are underlyingly stop-initial require a preceding /ŋu/ where they follow a bound stem, as in Example (4), where the verb follows an incorporated noun. The approximately 1% of agreement prefixes that don’t end in a vowel also condition insertion of /ŋu/, along with any other non-agreement morphology that precedes the verb, regardless of the final segment.
(3)
/ŋani-kucukucaani/
3masc/1sg-head-Ø-tickle. past.cont
‘He tickled me’
(4)
/ŋani-jinak-ŋu-kucukucaani/
3masc/1sg-head-Ø-tickle.past.cont
‘He tickled me on the head’
Morphophonological phenomena such as initial hardening and ŋu-insertion are typologically rare in the world’s languages but serve as excellent phenomena through which to study the processing of complex morphophonology. Indeed, such phenomena pose a challenge to models of processing because assumptions of input fidelity—i.e., that the input matches and therefore activates forms in the lexicon—cannot necessarily be satisfied. This is especially the case for changes to the initial segments, which have been demonstrated to be particularly important to lexical access (Cutler, 2008; Cutler et al., 1985).
The central questions of our research are: How do speakers process the related phenomena of phonological hardening and ŋu-insertion, and what theories of morphology—and the processing of morphology—can best account for the behavior of speakers? Fundamental to this pursuit is the question of whether complex words are stored as wholes in speakers’ mental lexicons or stored in partially or completely analysed forms (i.e., as morphemes or morpheme combinations) with some degree of computation involved in both comprehension and production (Amenta & Crepaldi, 2012; Marslen-Wilson, 2007; Zwitserlood, 2018). Or, can we allow for both storage and computation to varying degrees, as in dual-route models (Frauenfelder & Schreuder, 1991; Ullman et al., 1997; Ullman, 2001, 2004) based on factors such as productivity or regularity?
The predictions of these theoretical approaches differ, and studies like the one presented here may contribute to the assessment of their explanatory and predictive power. In full-listing approaches, only words have lexical entries (not sub-lexical morphemes), and so every complex word must be accessed by finding a matching entry in the mental lexicon (Baayen et al., 2011; Gonnerman et al., 2007; Plaut, 2011; Rueckl et al., 1997; Rumelhart & McClelland, 1986; Westermann & Ruh, 2012). These models typically assume an associative-connectionist network to model lexical access and retrieval. It is unclear how such approaches would deal with deviations from grammatical output such as mispronunciations or failures to implement a phonological rule, unless there is a separate mechanism to license such deviations from the canonical (underlying) form. As we have argued elsewhere (Bundgaard-Nielsen et al, 2025, p. 4):
Few would argue with the fact that [khæt] ‘cat’ is stored in the minds of English-speakers as /kæt/ and will not necessarily be well-recognised if produced as [ɡæt] (a change in the voicing specifications of the initial segment that crosses the phonetic VOT boundary) or [pæt] (a change in place of articulation), or [ɣæt] or [ɰæt] (a change in manner of articulation from a stop to a fricative [ɣ] or approximant [ɰ] at the velar place of articulation) by speakers. Changes of this sort potentially lead to activation of the wrong set of lexical competitors where the phonetic realisation results in a phonological substitution (/kæt/ ‘cat’ to /pæt/ ‘pat’ or /ɡæt/ ‘gat’, or no activation at all). These ‘wrong paths’ would require reinterpretation or repair of the input at some later point.
This is something that poses a significant challenge to models of processing, including associative and full listing models, that adopt a strictly linear approach.
In contrast, at least in some full-parsing and dual-route approaches (Chialant & Caramazza, 1995; Clahsen, 1999; Marslen-Wilson & Warren, 1994; Pinker & Ullman 2002; Schreuder & Baayen 1995; Taft, 2004; Ullman, 2001, 2004; Ullman et al., 1997), words are retrieved by parsing out morphemes and accessing lexical entries associated with those morphemes, including complex words containing them, at least in the case of productive/regularly formed words. Under these views, deviations from canonical forms can either be reverse engineered to reveal an existing underlying form—incurring a processing cost—or, in the case of dual-route models, result in direct activation of parallel listed forms in the lexicon. Mispronunciations that cannot be reverse engineered—such as the /pæt/ ‘pat’ or /ɡæt/ ‘gat’/ example from Bundgaard-Nielsen et al. (2025)—will either be repaired or rejected as unrecoverable (nonsense). See Blanco-Elorrieta et al., (2021) for discussion and evidence that repair of mispronunciation requires top-down processing. It is precisely these kinds of outputs that are examined here.
To address these questions, we examine speakers’ tolerance for deviations in the implementation of phonological hardening and ŋu-insertion, and how context affects their judgements. We do so in two Two Alternate Forced Choice (2AFC) experiments testing native speaker processing of Wubuy phonological hardening and ŋu-insertion. We show that Wubuy listeners do not simply accept all and only the grammatical outputs, and reject all the ungrammatical ones, which we might expect based on full listing models. Rather, they have variable rates of acceptance even of fully grammatical forms depending on the kind of grammatical processes and violations involved. The data also suggests that some forms are more costly or difficult to parse, suggestive of a computational process, in a manner inconsistent with full listing.
Our explanation for this behavior relies on the proposal that speakers store not all surface forms of morphemes in particular contexts (as in the full listing approach) but rather store “canonical” or “underlying” forms, from which all surface forms can be derived by the application of general processes, as proposed in the generative model of phonology in Chomsky and Halle (1968) and many works since (e.g., Lahiri & Marslen-Wilson, 1991; Gaskell & Marslen-Wilson, 1996; 1998; Gow, 2001; 2002, for similar approaches). The results reported here have implications for models of the lexicons of polysynthetic languages, as well as for the evaluation of theories of phonology, morphology, and language processing.
1.1. Morphophonological alternation in Wubuy
Wubuy (Heath, 1984) is a Non-Pama-Nyungan language spoken in a remote region of eastern Arnhem Land in the Northern Territory of Australia. It is related to other languages of Arnhem Land with larger speaker populations, such as Anindilyakwa (Van Egmond, 2012) and Bininj Gun-wok (Evans, 2003), and others that are effectively moribund, such as Dalabon (Evans, Fletcher, & Ross, 2008), Ngandi (Heath, 1978), and Ngalakgan (Baker, 2008; Merlan, 1983).
We estimate that Wubuy is spoken by around 60 fluent first language (L1) speakers primarily in the remote settlement of Numbulwar on the Gulf of Carpentaria (Bohn & Bundgaard-Nielsen, 2021). It is likely also an L2/L3 for some speakers in adjacent communities in north-east Arnhem Land, but the precise number of speakers is difficult to assess. Wubuy is primarily spoken by older people; children and young people in Numbulwar speak Kriol (an English-lexified creole: see Sandefur, 1982; Bundgaard-Nielsen & Baker, 2016), although they are often addressed in Wubuy by older members of the community and appear to have (some) receptive language skills.
Wubuy has a typical Australian phonological inventory (see Table 1), with four contrastive coronal places of articulation in stops and nasals, plus a dorso-velar and bilabial series. All four coronal series contrast word-initially as well as medially (which is also highly unusual, cross-linguistically, in Australia: Baker & Harvey, 2026, p. 74). Wubuy lacks voicing or constriction contrasts in obstruents; the obstruents are represented with the voiceless IPA series here, although their phonetic voicing and constriction values vary.
Table 1: Consonant inventory of Wubuy (after Heath, 1984).
|
Labial |
Lamino-dental |
Apico-alveolar |
Apico-retroflex |
Lamino-alveopalatal |
Dorso-velar |
| Stop |
p |
t̪ |
t |
ʈ |
c |
k |
| Nasal |
m |
n̪ |
n |
ɳ |
ɲ |
ŋ |
| Lateral |
|
l̪ |
l |
ɭ |
|
|
| Trill/tap |
|
|
r |
|
|
|
| Approximant |
w |
|
|
ɻ |
j |
|
Wubuy is characterised by extensive morphophonological alternations. Due to a historical process of chain shift lenition (Heath 1978; 1984), Wubuy morphemes overwhelmingly begin in continuants, including vowels. A subset of the phonemic continuants {w, j, ɻ, r, l̪} alternate in a predictable fashion with place-matched stops following non-continuants (nasals and stops) in a process dubbed hardening in Heath (1984); see Example (1) and Example (2), where the verb-initial continuant /w/ alternates with /k/ when a nasal or stop precedes (other w-initial morphemes alternate with /p/: the difference is lexically specified and unpredictable). Phonological hardening is not a minor phenomenon in Wubuy: ~57% of all roots in the lexicon of Heath (1982) begin in one of these five alternating continuants (Bundgaard-Nielsen & Baker, 2014), more than half of them in /w/. Phonological hardening is completely productive and exceptionless in Wubuy: All morphemes, including verbs, which begin in {w, j, ɻ, r, l̪}, undergo this hardening in the relevant inter-morphemic environment.1
Phonological hardening operates in parallel with a second process—ŋu-insertion—in which the meaningless string /ŋu/ is inserted before underlyingly (i.e., not derived from a continuant by hardening) stop-initial morphemes, when they occur in environments which partially overlap the environment for hardening, as in Example (4), presented above. Here, underlying stop-initial verbs such as kucukucaani ‘tickle.past.cont’ must have a preceding string ŋu if either of the following conditions is met:
(a) the preceding morpheme is one of the only three agreement prefixes ending in a non-continuant (out of the approximately 350 agreement prefix forms), or
(b) the preceding morpheme is a lexical stem of any phonological shape.
In Example (4), the final stop of the preceding morpheme jinak ‘head’ is (optionally but frequently) deleted in production, but we infer its presence because of the behavior of verbs that begin in continuants: these occur in their hardened, stop-initial forms following jinak and similar incorporable nouns ending in stops, regardless of stop realization. Heath (1984; and compare Nyhuis, 2023) argues that ŋu-insertion thus serves as an indication that the following morpheme is underlyingly stop-initial, because without it, a phonetic morpheme-initial stop could be underlying or derived, following a non-continuant in the previous morpheme, as in (4). Accounting for the parsing of this semantically empty element is difficult within any linear framework, including associative or listing models of processing, as the need to insert /ŋu-/ only arises once the lexical specifications—or at a minimum, the onset—of the following morpheme can be known, requiring a degree of “look-ahead” commonly ruled out in morphological theories (e.g., Simpson & Withgott, 1986, p. 155; Carstairs-McCarthy, 1987).2
1.2 Competing theories of word formation
Morphophonological patterns such as Wubuy phonological hardening and ŋu-insertion challenge several models of word-formation. Under associative models of morphology like Construction Morphology (e.g., Booij, 2018; Nyhuis, 2023) and Relational Morphology (Jackendoff & Audring, 2020)—as well as under linear and associative models of word recognition (e.g., the Cohort Model: Marslen-Wilson & Welsh, 1978; TRACE: McClelland & Elman, 1986; Shortlist B: Norris & McQueen, 2008)—language acquisition is “item based.” This means that “the composition of a word or sentence involves clipping together stored pieces [words or morphemes, depending on the model] in such a way that every element of the composed structure is accounted for in terms of one stored piece or another” (Jackendoff & Audring, 2020, p. 4; our insertion in brackets). Lexico-grammatical choices are thus explained primarily by complex patterns of parallel associations between (adjacent) elements. These elements form independent systems, connected by linkages that create “correspondences between components rather than derivations from one component to another” (Jackendoff & Audring, 2020, p. 2).
Under associative and similar full-listing models, phonologically conditioned changes or alternations such as those in Wubuy must be accounted for as non-derivational and parallel, and the involved morphemes/words must be selected in a one-step process from multiple equivalent forms in the lexicon, based on their co-occurrence patterns in the input. While frequency in the input might affect the strength of the linkages between co-occurring elements, such that infrequent sequences are less accurate or even slower, for frequently occurring sequences, these models would expect speakers/hearers to select the appropriate grammatical sequence of morphemes with ease and high levels of accuracy. This is what we would expect in Wubuy, given that both hardening and ŋu-insertion are highly frequent and obligatory (see Section 1.1). Note that despite talking of words here, Wubuy verbs are complete propositions, translating into entire sentences in English as per the first four examples, and their processing shares aspects of the processing of sentences in some respects. They can also have the prosodic structure typical of phrases, including internal major junctures characteristic of Intonational Phrase boundaries (for a discussion, see Bundgaard-Nielsen & Baker, 2020).
Under competing rule-based analyses, building on Chomsky and Halle (1968), individual morphemes have one canonical phonological shape in the lexicon, which must be selected prior to the environment triggering the application of any (morpho-)phonological rules (Levelt, 1989; Levelt et al., 1999). This is a sequential process (even if constraints on surface forms are computed in parallel, as in Optimality Theory: Prince & Smolensky, 1993), and the identification and implementation of necessary phonologically conditioned changes presumably increase the processing demands of spoken word recognition and comprehension when a phonologically conditioned change must be applied. We are aware of no work on this specific issue, but see Bentum et al. (2019), Gaskell & Marslen-Wilson (1996; 1998) and, more recently, Stockall & Gwilliams (2023). A rich literature has examined the processing (recognition) of words produced in canonical—correct, clear, unreduced—phonetic form relative to phonetically reduced or mispronounced forms. Such research generally shows better recognition for canonical forms, though non-canonical word recognition improves when contextual cues are strong (see e.g., Brouwer et al., 2013) and when non-canonical forms are highly frequent (see e.g., Bürki et al., 2018).
In the case of the Wubuy hardening phenomenon, we would expect application of a hardening rule to induce greater processing cost than non-application of a rule (no change, as the trigger for change is absent) and likely higher error rates and greater time cost associated with rule application. Such asymmetrical performance is at odds with the predictions from full listing models, especially if stark differences in the frequency of forms are not apparent. Other models offer more flexible processing paths. The Dual-Route Hypothesis (Frauenfelder & Schreuder, 1991; Ullman et al., 1997; Ullman, 2001, 2004) suggests that regular (predictable) and irregular (unpredictable) patterns might be processed in two different ways. While rule-governed regular patterns are processed implicitly, irregular patterns are retrieved from declarative memory.
Since most of the processing literature has focused on a very narrow typological range of languages, there is limited literature on similar or related types of processes affecting the initial segments of morphemes. Initial consonant mutation in Celtic languages, such as Welsh and Gaelic, and word-initial mutation and lenition in the Australian language Mawng, are two of the rare counterexamples. In Celtic languages such as Welsh, the initial consonants of nouns alternate depending on their grammatical environment, e.g., the word /pont/ ‘bridge’ may appear also as /bont/ or /font/ (Ball & Müller, 1992; Fife & King, 1998). But Celtic mutations cannot be predicted based on the synchronic phonology (i.e., in terms of the local segments in the environment). Rather, the conditioning is dependent on syntactic factors like ‘definite NP’ or ‘genitive NP’. Hence, processing accounts of Celtic mutation rely on accessing items in memory (Ussishkin et al., 2017).
Mawng is also described as having lexicalised initial mutations in verb and noun roots (Evans, 2000, p. 102), as shown in Example (5) in bold. These mutations could presumably be accounted for in a similar way to Welsh mutations. Mawng, however, is also described as having synchronic lenition processes affecting initial segments at word boundaries (Capell & Hinch, 1970, p. 44), as exemplified in Example (6). Example (6)a and Example (6)c show that initial /ɡ/ can vary between stop and approximant realisations, following a vowel in the preceding word. This variation is not licenced/observed following a non-continuant, as in Example (6)b and Example (6)d.
(5)
a.
a-mawur
3pl-arm
‘their arms’
b.
bawur
3.arm
‘his arm’
(6)
a.
ɡe
it.goes
[ɡ ~ ɣ]apala
boat
‘the boat goes’
b.
ɡaɾɡbin
big
[ɡ]apala /*ɣabala
boat
‘big boat’
c.
mada
veg.the
[ɡ ~ w]ubuɲ
canoe
‘the canoe’
d.
mariɡ
NEG
[ɡ]ubuɲ/*wubuɲ
canoe
‘not a canoe’
In contrast with Celtic languages, the processing of lenited onsets in Mawng invokes accounts based in the phonetics/phonology/lexicon interface (Bundgaard-Nielsen et al., 2025) and has been argued to be context-conditioned (Capell & Hinch, 1970; Shaw et al., 2020). Together, this suggests that different languages may rely on different processing strategies: some patterns may best be accounted for by associative patterns, while others are best accounted for through rule application. Other languages may require speakers to use both approaches and handle some patterns via associative patterns and expansive listing and others via rule application, as suggested, for instance, by the Dual-Route Hypothesis (Chialant & Caramazza, 1995; Clahsen, 1999; Marslen-Wilson & Warren, 1994; Pinker & Ullman 2002; Schreuder & Baayen 1995; Taft, 2004; Ullman et al., 1997; Ullman, 2001, 2004), and applied by Baker (2008) in an analysis of the Australian language Ngalakgan.
1.3 The present studies
In what follows, we test the effect of phonological hardening and ŋu-insertion on word processing in two Two-Alternate Forced Choice (2AFC) experiments with speakers of Wubuy. Experiment 1 (Section 2) tests the effect of over- and under-application of phonological hardening on word recognition in a two-part experiment which differs only in whether a short 500 ms pause has been inserted before the hardened target verb root. This pause provides participants with additional time to process the frame and thus identify the need for hardening. Since hardening is such a frequent process in Wubuy (see Section 1), we assume that speakers accord it a significant degree of attention. The results indicate that Wubuy speakers find it easier to select the target form of a verb if it does not require hardening, and accuracy is enhanced when a pause is artificially inserted before the target verb to allow parsing of the conditioning context. The existence of a mid-word pause does not cause a naturalness problem: word-internal pauses occur frequently in unmonitored speech in Wubuy and are accepted—indeed, even preferred in some cases—by speakers in a previous 2AFC experiment (Bundgaard-Nielsen & Baker, 2020), even where a pause separates the environment for hardening from its target as in /ŋa-Ø-ɻaŋa- … kulta̪ŋi/ (1sg-resid-branch-cut.past.cont), where ‘…’ indicates inserted pause and kulta̪ŋi is the same hardened verb as in (2), hardened because of preceding noun /ɻaŋak/ ‘branch’ where the final /k/ has been deleted in the produced form. The results are largely consistent with generative/rule-governed models, though the processing of some lexemes might rely on listing or memory-based approaches (see Section 2.2.1).
Experiment 2a (Section 3.2) tests ŋu-insertion preference with verb forms that vary in their initial segment characteristics—stop-initial, continuant-initial, hardened and stop mispronounced as continuant. Experiment 2b (Section 3.3) tests verb form preference in frames with or without ŋu-insertion. The results indicate that Wubuy speakers prefer grammatical utterances in both experiments and that grammatical errors and mispronunciations disrupt processing in gradient ways that are, again, not straightforwardly compatible with linear accounts of word recognition, such as full listing models.
Together, the studies suggest that language phenomena like the ones examined here—requiring speakers/listeners to process sequences of morphemes that change shape in ways that are not forwardly predictable—incur a processing cost that is hard to explain in approaches relying on linear and probabilistic associations or connections (Zwitserlood, 2018). Rule-based accounts may have more success accounting for such patterns, and the results presented here do indicate that correct parsing of this kind of language material is costly for the listener in ways consistent with rule-based assumptions.
However, before we provide the details of Experiment 1 and Experiment 2, we wish to preface these sections with additional background. In many Indigenous Australian communities, death taboos are observed that prevent individuals from naming, looking at images, photos and videos of, and/or listening to recordings of other Indigenous people who have passed away, particularly people who are well-known to them, e.g., close family or other community members. This is relevant here, because, very sadly, the speaker who produced the stimulus materials for Experiment 1 and Experiment 2 passed away during the data collection phase (data was collected during several field trips). Consequently, we are unable to collect further data using these materials for cultural reasons. The studies were approved by a Human Ethics Panel at the University of Melbourne (#26884) and funded by the Australian Research Council (DP180100821).
4. General discussion
The present paper examines the processing of two linked morphophonological alternations—phonological hardening and ŋu-insertion—in the Australian Indigenous language Wubuy. The studies probe the crucial theoretical question of whether such linguistic phenomena are better conceptualised as morpho-lexical parallel-associative processes (e.g., Construction Morphology: Booij, 2018; Relational Morphology: Jackendoff & Audring, 2020), should instead be accounted for by rule-governed processes on underlying forms (e.g., Chomsky & Halle, 1968), or perhaps whether both memory- and rule-based processing (Ullman et al., 1997; Ullman, 2001; 2004) are necessary to account for speaker behavior.
Experiment 1 (Section 2) tested the effect of over- and under-application of phonological hardening on word recognition in a two-part 2AFC experiment (Experiment 1a, 1b). Each trial required participants to indicate preference between word pairs where hardening was correctly applied in half, and under-applied or over-applied (i.e., with/without a preceding non-continuant segment, respectively) in the other half. Experiment 1a and Experiment 1b differed only in whether a brief pause (500 ms) had been inserted before the target verb. The results, excepting the special case of -aajuu and its stop-initial counterpart -kaajuu, indicate that it is easier for Wubuy speakers to correctly select non-hardened verbs in non-hardening contexts than it is for them to correctly select hardened verbs in a hardening context, and that extra processing time (here, pause insertion) improves accuracy, in line with the rule-based predictions in Section 2.1.4. Overall, this pattern of behavior—increased difficulty in one direction and improvement in accuracy due to pause insertion—is consistent with the predictions of rule-based approaches. It suggests that the lexicon only stores canonical (non-hardened) forms, with the exception, perhaps, of vowel-initial verbs—in this case, the insertion of an onset consonant attains the correct surface form. The results presented here are more consistent with an account where the correct form is selected using declarative memory.
Experiment 2 (Section 3) examined the process of ŋu-insertion—the insertion of the meaningless string /ŋu-/ before morphemes beginning in an underlying stop—under complex conditions but including where the preceding morpheme ends in an underlying or surface stop, as is the case with the complex transitive Benefactive prefix nuŋaa-. Experiment 2a (Section 3.2) tested the preference for ŋu-insertion with four types of target verbs: verbs beginning in faithful stops, verbs beginning in faithful continuants, verbs beginning in phonologically hardened stops, and mispronounced, pseudo-lenited verbs.
The results from Experiment 2a (unsurprisingly) demonstrate that participants prefer grammatical sequences, as would be expected under parallel-associative and rule-governed approaches, but also that different types of errors induce different levels of rejection, which, we argue, reflect participant knowledge of multilevel processes of word formation. A clear example of this is that preference patterns differ for the two verb forms (the continuant-initial verbs and the mispronounced verbs) that are ungrammatical with both prefix frames. While participants clearly prefer the frame that would be grammatical for the mispronounced form if it were pronounced with an initial stop (76% preference for the nuŋaaŋu- frame), frame preference for the continuant-initial form is at chance, even though ignoring the continuancy of the initial segment (and pretending it is a stop, as with the mispronounced form) would give a grammatical result in the nuŋaa- frame). This suggests that listeners prefer to ignore a single-feature mispronunciation to ensure grammaticality for the mispronounced forms (preferring mispronounced forms in the correct nuŋaaŋu- frame), because by doing so, they can arrive at a grammatical form (with an underlying stop). By contrast, participants are not willing to ignore a single-feature mispronunciation in the form of an unhardened continuant-initial form in the nuŋaa- frame (it is deemed a no better choice than the double-error option offered by the nuŋaaŋu- frame). This suggests the existence of two different processes in dealing with these inputs (continuant-initial and mispronounced). In the continuant-initial input, speakers recognize the verb morpheme, but neither frame produces a grammatical result, giving them no good options. In the mispronounced input, speakers fail to find a match with an existing morpheme, and no morphophonological rule is available to transform the input into an underlying form, which perhaps licenses their inference that it is mispronounced, but not necessarily ungrammatical.
Experiment 2b (Section 3.3) tested the preference for stop and continuant-initial verb onsets (hardened stops versus underlying continuant-initial verbs; stop-initial versus unattested mispronounced verbs) in both the nuŋaa- and nuŋaaŋu- frames. The results show—again unsurprisingly—that the Wubuy participants prefer inputs in the correct grammatical form, but the results also indicate that the participants are willing to accept canonical verb forms when they are presented in an ungrammatical context. This latter result indicates that the preceding context does not preclude preference for verbs whose shape is not licenced by the context when the alternative choice—despite having a suitable form—cannot be converted (by rule application or other mechanism) into a meaningful selection from the lexicon. Therefore, we cautiously suggest that parallel-associative approaches relying on a listener “clipping together stored pieces in such a way that every element of the composed structure is accounted for in terms of one stored piece or another” (Jackendoff and Audring, 2020, p. 4; our emphasis) may not adequately account for the observed behavior.
In their totality, the results from Experiment 1 (hardening) and Experiment 2 (ŋu-insertion) presented here suggest that the twin phenomena of phonological hardening and ŋu-insertion in the Australian Indigenous language Wubuy (Heath, 1984) are better accounted for as multi-step, rule-based processes, as the results are consistent with a Wubuy lexicon where verbs are listed in a single form. For continuant-initial verbs, this means they do not have a stop-onset parallel listed form that is activated when words and phrases are, to paraphrase Jackendoff & Audring (2020), “clipped together” in a manner that accounts for all elements and on the basis of frequency and distributional patterns. Rather, as illustrated by the results of Experiment 1, that hardening is the result of a rule-based process, one that increases the cognitive load and increases the error rate.
For stop-initial verbs, only a single lexical entry is assumed and preference for the grammatically correct nuŋaaŋu- frame is high and very consistent in Experiment 2a. Participants also confidently prefer the grammatically correct nuŋaaŋu- frame (rather than incorrect nuŋaa-) even when stop-initial verbs are produced with a single-feature onset mispronunciation as a continuant. Participants’ equally strong preference for correctly produced stop-onset verbs over mispronounced forms in Experiment 2b demonstrates that mispronunciation tolerance is not due to stop-continuant phonological neutralisation following /ŋu-/, though introduction of a grammatical error (undersupplied /ŋu-/), also in Experiment 2b, shows that increased processing demands in the form of a grammatical mismatch increases participant tolerance for the mispronounced forms (a reduction in accuracy from 88% to 74%).
Under parallel-associative models of morphology, as well as linear models of continuous parsing like the Cohort Model (Marslen-Wilson & Welsh, 1978), TRACE (McClelland & Elman, 1986) and Shortlist B (Norris & McQueen, 2008), the effect of frame on verb preference is important, as correct selection of upcoming elements from parallel listings in the lexicon is contingent on what has already been presented. Under these models, nuŋaa- is presumably processed first. A listener then either:
(1) identifies /ŋu-/ and then proceeds to select a verb form from those verbs in the lexicon that have only one (stop-initial) entry listed, or,
(2) proceeds directly to select a stop-initial verb form from those lexical entries that have two parallel listed forms, one of which is stop-initial (hardened) and one of which is continuant/vowel initial.
Given that the requirement of ŋu- is not determined by the prefix alone but by the prefix in combination with the onset of the verb, it is not possible for speakers to determine whether /ŋu-/ is required, however, until verb-selection is complete. This challenges the linearity of the process, unless one wishes to argue that all stop-initial verbs are also dual-listed, with a form starting in /ŋu-/ (though see Nyhuis, 2023 p. 106, for an alternative position).
Rather, the data presented here is consistent with an approach to word recognition in which phonological hardening and ŋu-insertion are processes that occur after all lexemes have been selected, as in Levelt’s classic processing model (1989). Under this approach, prefix and verb selection precede the application of any morphophonological rules (hardening or ŋu-insertion), which are triggered only by the phonological form of the selected item, once the item has been selected. Where a rule is triggered, it activates a secondary process, incurring additional processing costs, which presumably can be quantified in the form of reaction time and accuracy measures, as demonstrated here.
The data presented here does not, however, suggest that all processes must be rule governed, and that parallel listings and associative patterns cannot account for some aspects of language processing, including some aspects of the complex Wubuy morphophonology that we are discussing here. Indeed, some of the data presented here, as in the case for the alternations in vowel-initial verbs like -aajuu, is better accounted for by exactly the kind of memory and listings-based approach advocated for in parallel-associative models, and we encourage consideration of models of word recognition that allow associative/listings-based explanations of some phenomena and rule-based explanations for others in an integrated fashion (Ullman et al., 1997; Ullman, 2001, 2004). However, the results also suggest that parallel-associative and Constructionist approaches would need additional architecture to explain why (a) some grammatical forms are apparently more difficult to select than others, and (b) why some unattested forms are better than others.
Finally, we wish to address the limitations of the present experiments. Experimental research on language processing in Wubuy is constrained by many factors, including language loss, an aging population, and substantial logistical difficulties associated with data collection in remote parts of Australia. The effect of these constraints warrant discussion. First, our experimental designs and approaches consequently deviate from standard laboratory practices, as many traditional laboratory approaches are unsuitable for participants who require technical support and prefer to be tested in smaller social groupings. This means that the data collected is less controlled than is typically the case and that certain types of data that might have been highly informative (e.g., response times) cannot be collected meaningfully. Second, we acknowledge that the low number of participants, as well as the low number of experimental trials per participant, dramatically reduces the statistical power of our experiments, with consequences for the generalizability of the results. We have detailed the reasons for this elsewhere, and we would like to highlight again, that Experiment 1 tested 25% of the speaker population, while Experiment 2 tested approximately 15%. This is a very significant coverage of the population, akin to testing millions of speakers of Australian English in a study of the processing (or other phenomena) of that variety of English. We would also like to highlight that research using the experimental protocol employed here, along with similar participant numbers, has already contributed substantially to our understanding of language processing in Australian Indigenous languages (Bundgaard-Nielsen & Baker, 2020; Bundgaard-Nielsen et al., 2025). And in pursuit of hard-to-get experimental data, we encourage colleagues to employ whatever design is acceptable to the speaker population. More research of this kind will greatly benefit the research community and theory building, including theories of continuous parsing, as recognized in several recent overviews (Anand et al., 2011; Berghoff & Bylund, 2025; Norcliffe et al., 2015).